This article presents a methodology to estimate the live fuel moisture content (LFMC), a critical factor in the spread of forest fires, through machine learning tools. Random forest models were trained using field LFMC samples collected bi-weekly for 18 consecutive months in 43 shrubland plots in the Valencian region, a Mediterranean zone in eastern Spain. LFMC predictions were obtained for the weighted average of LFMC values, calculated using the Fraction of Canopy Cover (FCC) of dominant species as weights. Furthermore, a specific model was defined for predicting LFMC for the Rosmarinus officinalis species. A Forward Feature Selection (FFS) with a Leave-Location-Out Cross Validation (LLOCV) method was used to select predictors extracted from a spatiotemporal data set, which includes different spectral indices obtained from Sentinel-2 imagery and meteorological variables obtained from measurements at weather stations, along with other seasonal, geographical or topographic variables. Model predictions were validated with a LLOCV procedure, and also using independent field measurements of LFMC in another period with changes in the precipitation regime and average temperatures. Variables selected by FFS for the two LFMC models were: the cumulative precipitation in the previous 60 days (p60), the average of the daily mean temperature in the previous 60 days (t60), together with the Y-UTM coordinate and the sine and cosine of the day of the year. LFMC predictions for the weighted average of LFMC values also introduced the Transformed Chlorophyll Absorption Ratio Index (TCARI), resulting in an R2 of 68.1 %. However, LFMC for the Rosmarinus officinalis species used the ratio between TCARI and the Optimized Soil-Adjusted Vegetation Index (OSAVI), in addition to the average daily minimum relative humidity in the 15 days prior to the date considered (R2 = 74.9 %). LFMC time series analysis showed that the general trend of LFMC measures is satisfactorily captured by the predictions. Spatial and temporal variations in LFMC were analyzed throughout thematic maps in the studied area during the wildfire season.
Read full abstract